Professional Documents
Culture Documents
Ckasumo MSC Dissertation
Ckasumo MSC Dissertation
net/publication/311424056
CITATIONS READS
6 701
1 author:
Christian Kasumo
Mulungushi University
22 PUBLICATIONS 0 CITATIONS
SEE PROFILE
All content following this page was uploaded by Christian Kasumo on 05 December 2016.
Christian Kasumo
September 2011
MINIMIZING THE PROBABILITY OF ULTIMATE
RUIN BY PROPORTIONAL REINSURANCE AND
INVESTMENTS
By
Christian Kasumo
CERTIFICATION
The undersigned certify that they have read and hereby recommend for acceptance
by the University of Dar es Salaam a dissertation entitled Minimizing the Prob-
ability of Ultimate Ruin by Proportional Reinsurance and Investments,
(Supervisor)
Date:
(Supervisor)
Date:
ii
I, Christian Kasumo, declare that this dissertation is my own original work and
that it has not been presented and will not be presented to any other University for
a similar or any other degree award.
Signature:
This dissertation is copyright material protected under the Berne Convention, the
Copyright Act 1999 and other international and national enactments, in that behalf,
on intellectual property. It may not be reproduced by any means, in full or in part,
except for short extracts in fair dealings, for research or private study, critical schol-
ACKNOWLEDGEMENTS
The author wishes to pay tribute to several persons without whose help, direct or
indirect, this work would not have been completed. First and foremost, I wish to
express my gratitude to God for giving me such a wonderful opportunity to study in
Tanzania and for supplying the understanding I needed to pursue my research and
complete my studies. Without Him, I could have done absolutely nothing.
(UDSM, Tanzania) and Dr Juma Kasozi (Makerere University, Uganda) who saw
to it that I made adequate progress in my research and offered valuable corrective
suggestions and guidance throughout. I will forever value the encouragement and
support that they gave me throughout the hectic period of research and dissertation
writing. I cannot find words to adequately thank them for all their help. Through
their knowledge and expertise I have learned much about insurance modelling and
without them this document would simply not have been completed.
In the fourth place, I wish, from the bottom of my heart, to thank the Council
and Management of Mulungushi University (MU) for the scholarship awarded to me
Fifthly, let me take this opportunity to thank the Coordinator of the Norad Pro-
UDSM. They were there for me when the going got tough and through their friend-
ship I was able to safely navigate the stormy seas of scientific research and to reach
the shore.
Finally, my heartfelt thanks go to my dear wife, Baptista, and our daughters, Chab-
ota, Choolwe, Chimuka and Christabel, who put up with my long absence from home
so that I might pursue my postgraduate studies in Dar es Salaam. Even when I was
home and busy working on my research, they were very understanding and gave me
the time and space I needed to complete my research. I am blessed to have such a
wonderful family.
v
DEDICATION
To my family
Baptista, Chabota, Choolwe, Chimuka and Christabel
vi
ABSTRACT
The study was conducted on the topic: Minimizing the Probability of Ultimate Ruin
by Proportional Reinsurance and Investments. The purpose of the study was to
determine the role of investments in minimizing the probability of ultimate ruin of an
tion of the second kind. This integral equation was then solved using the block-by-
block numerical method for the retention percentage that minimizes the probability
of ultimate ruin. The major findings of this study were as follows:
1. That, as expected, the higher the investment rate, the lower the ruin probabil-
ity. Furthermore, the study has revealed that volatility of stock prices results
2. That for a given initial surplus, the ruin probabilities keep reducing as the value
of the retention level b reduces. After a certain b, however, the ruin probabilities
begin rising again, giving an indication of the location of the optimal retention
percentage b∗ .
3. That the optimal retention level, given certain assumptions regarding the flow
of premium incomes, is b∗ = 0.315034 for the small claim case and b∗ = 0.461538
in the case of large claims.
vii
Some recommendations have also been made with regard to strategies that could be
used by insurance companies to minimize their ultimate ruin probabilities. The study
has recommended that in order to minimize their ruin probabilities insurers should
invest their surplus in both risky and risk-free assets. It has also been recommended
that insurers buy reinsurance as it helps in reducing the probability of ultimate
ruin for insurance companies. But, given certain assumptions regarding the flow of
premium incomes, insurers can only reinsure optimally when b∗ = 0.315034 for small
claims and b∗ = 0.461538 for large ones.
Contents
Certification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . i
Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii
Dedication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v
Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi
Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xv
1 CHAPTER ONE
INTRODUCTION 1
viii
ix
1.6 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2 CHAPTER TWO
LITERATURE REVIEW 18
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.2 Reinsurance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.3 Investments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3 CHAPTER THREE
MODEL FORMULATION 27
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4 CHAPTER FOUR
MODEL ANALYSIS 35
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
4.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
5 CHAPTER FIVE
NUMERICAL RESULTS 44
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
5.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
xi
5.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
6 CHAPTER SIX
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
6.2 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
6.3 Recommendations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
REFERENCES 67
APPENDICES 75
List of Tables
µ = 0.5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
5.5 Ruin probabilities for C-L model with proportional reinsurance for
p = 6, λ = 2, µ = 0.5, 0.2 ≤ b ≤ 1.0 and h = 0.01 . . . . . . . . . . . 56
2, α = 3 and κ = 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
2, α = 2 and κ = 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
5.9 Asymptotic ruin probabilities for large claims with proportional rein-
xii
List of Figures
5.1 Effect on ruin probability of changes in volatility of the stock price (σR ) 63
xiii
xiv
LIST OF ABBREVIATIONS
HJB Hamilton-Jacobi-Bellman
ECOMOR Excédent du coût moyen relatif (‘excess of the average cost’)
XL Excess-of-loss
B-by-B Block-by-block method
VIDE Volterra integro-differential equation
VIE Volterra integral equation
GBM Geometric Brownian Motion
BM Brownian Motion
SDE Stochastic differential equation
C-L Cramér-Lundberg model
iid independent and identically distributed
a.s. almost surely
iff if and only if
w.r.t. with respect to
s.t. such that
xv
NOTATION
This study seeks to investigate the impact of proportional reinsurance and invest-
ments on the ruin probability of an insurance company. The purpose of the study is
to determine the role of investments in minimizing the probability of ultimate ruin
key terms used in the study and gives a statement of the problem. It also states
the objectives of the study as well as the research hypotheses that will guide the
research. In addition, the chapter indicates the significance of the study and outlines
Modern life is characterized by risks of different kinds, some threatening all persons
and others restricted to the owners of property, while still others are typical for some
individuals or for special occupations. Insurance was born to take all or part of
such risks from the risk-bearer. Under an insurance contract the insurer accepts to
pay part or all of the policyholder’s loss on the occurrence of an uncertain specified
event or risk covered by the contract. In return for this protection or cover the
policyholder accepts to pay the prescribed sum of money, called a premium, on an
agreed regular basis into a pool out of which the unfortunate individuals who actually
1
2
Insurance therefore operates on the basis of the fortunate many (who do not suffer
loss) helping the unfortunate few (who actually suffer loss through the occurrence of
the risk concerned). In other words, insurance cover or protection is accomplished
through a pooling mechanism whereby many individuals who are vulnerable to a
common risk are joined together into a risk pool. By pooling both the financial
contributions and the risks of a large number of policyholders, the insurer is typically
able to absorb losses incurred over any given period much more easily than would
However, in providing insurance cover, insurance companies run the risk of their
surplus becoming negative, thus resulting in ‘technical ruin’ (Dutang et al., 2009).
Insurers must therefore take measures to reduce the chances of their surplus becoming
non-positive (that is, the probability of ruin). These measures include investment
of the surplus, taking reinsurance, portfolio selection and volume control through
the setting of premiums. Hipp (2004) referred to these possible actions as control
variables. This study focuses on reinsurance coupled with investment as mechanisms
for reducing the risk in an insurance portfolio. Reinsurance (which is the insurance
of insurance companies) is essential not only from a legal standpoint but also due to
the fact that the risk left to the first-line insurer is usually substantially large.
return for a payment called a premium (Encarta World English Dictionary, 1999).
3
Definition 1.1.2. (Reinsurance) This is the transfer of risk from a direct insurer
(the cedent) to a second insurance carrier (the reinsurer). It may also be defined
Definition 1.1.3. (Risk) This is the probability of loss to an insurer or the amount
that an insurer is in danger of losing (Microsoft Encarta Encyclopedia, 2009).
with the modeling of claims that arrive in an insurance business and which gives
advice on how much premium has to be charged in order to avoid bankruptcy (ruin)
of the insurance company’ (Mikosch, 2004).
Definition 1.1.5. (Claim) A claim is a demand made by the insured, or the in-
sured’s beneficiary, for payment of the benefits as provided by the insurance policy
(Insurance Dictionary, 2009).
With reinsurance, therefore, the first-line insurer passes on some of its premium
income to a reinsurer who, in turn, covers a certain proportion of the claims that
occur (Albrecher and Thonhauser, 2009). Indeed, it has been argued very specifically
by Engelmann and Kipp (1995) that reinsurance plays an important role in risk-
reduction for first-line insurers in that it is able to offer additional underwriting
capacity for cedents as well as to reduce the probability of a direct insurer’s ruin.
This is one of the hypotheses that is investigated in this study. But if, in addition,
the company decides to invest its surplus, it must decide on an investment strategy
that aims at contributing to the minimization of the infinite time ruin probability.
minimizes the probability of ultimate ruin for an insurance company in the presence
of investments.
Definition 1.1.6. (Surplus) A surplus is the amount of money that remains after
all liabilities have been met. In other words, an insurance company accumulates a
surplus when the premiums collected over a given period of time exceed the claims
that have been paid over that period (Encarta World English Dictionary, 1999; Mo,
2002).
The difference between insurance and risk theory highlighted in Definitions 1.1.1
and 1.1.4 is also underscored by Malinovskii (2000). According to him, insurance
is a method of coping with risks, while the object of risk theory is to give a mathe-
matical analysis of random fluctuations in the insurance business and to discuss the
various means of protection against their inconvenient effects. Classical risk theory
focuses its attention on the outflow process, looking first at claim numbers, then
at the distribution of claim sizes and finally combining these two into an aggregate
claim amount process. The income process, which is the initial capital plus premium
income, is introduced in a fairly simple way, growing linearly in time at a constant
rate. If the insurance company also invests some of its surplus, then the resulting
surplus process of the company can be thought of as:
now be defined:
5
Definition 1.1.7. (Sample space) A sample space Ω is the set of all possible
outcomes of some random experiment (Kijima, 2003).
1. ∅, Ω ∈ F
2. If A ⊂ Ω ∈ F, then Ac = Ω\ A ∈ F.
S∞ T∞
3. If A1 , A2 , . . . ∈ F, then k=1 Ak , k=1 Ak ∈ F
1. P(∅) = 0, P(Ω) = 1
S∞ P∞
3. If A1 , A2 , . . . ∈ F, then P( k=1 Ak ) ≤ k=1 P(Ak )
S∞ P∞
4. If A1 , A2 , ... is a sequence of disjoint sets in F, then P( k=1 Ak ) = k=1 P(Ak )
It follows that if A, B ∈ F, then A ⊆ B implies P(A) ≤ P(B)
Definition 1.1.10. (Random variable) Let Ω be a non-empty finite set and let
Definition 1.1.11. (Probability space) A Probability space is the triple (Ω, F, P),
where Ω is a non-empty set, F is a σ-algebra consisting of subsets of Ω and P the
mation up to time t ≥ 0 (that is, which determines the timing of the revelation of
information) (Irgens and Paulsen, 2004; Duffie, 2009).
It will be assumed that all stochastic quantities and random variables (see Defini-
tions 1.1.10 and 1.1.14) are defined on a large enough filtered probability space
(Ω, F, {F}t∈R+ , P) where Ω is the sample space, F the σ-algebra on Ω, {F}t∈R+ the
filtration and P the probability measure defined on F.
càdlàg (continue à droit, limite à gauche) sample paths. For a detailed definition,
see Geiss (2010), Mikosch (2004) and Ramasubramanian (2005).
counting process N = {N (t)}t∈R+ which models the number of claims that occurred
up to time t ≥ 0 (Mikosch, 2004).
The most common claim number process is the Poisson process because of its good
theoretical properties (as seen in Definition 1.1.15). The jump sizes (or claim sizes)
will be modelled by an exponential distribution (for small claims). In his 1903 the-
sis, Lundberg exploited it as a model for the claim number process {N (t)} (Mikosch,
2004). However, because the (homogeneous) Poisson process does not always de-
scribe claim arrivals in an adequate way, other processes to model the number of
claims were developed. Examples of these are:
1. Renewal process
T0 = 0, Tn = X1 + X2 + ... + Xn , n ≥ 1,
N (t) = Ñ (θµ(t)), t ≥ 0,
8
between a homogeneous Poisson process and an inhomogeneous one is that for the
former the arrival rate λ is constant, whereas the latter allows for the arrival rate to
be a function of time λ(t) (Daniel, 2007; Gershman and Wilkerson, 2010).
Mikosch (2004) further asserts that a homogeneous Poisson process {N (t)}t∈R+ with
intensity λ:
2. starts at zero,
3. has independent and stationary increments, that is, for any 0 ≤ s ≤ t and
h > 0, N (s, t) has the same distribution as N (s + h, t + h). That is, the random
variables N (t) − N (s) and N (t + h) − N (s + h) have the same distribution
or probability law. This means that the probability law of the number of
claim arrivals in any interval of time depends on the length of the interval
(Ramasubramanian, 2005),
ramanian, 2005).
9
The homogeneous Poisson process is one of the prime examples of Lévy processes
with vast applications in life. Ramasubramanian (2005) adds two other properties
of the claim number process {N (t)}:
1. The probability of two or more claim arrivals in a very short time span is
negligible, that is, P(N (h) ≥ 2) = o(h), as h ↓ 0.
2. In a very short time interval, the probability of exactly one claim arrival
is roughly proportional to the length of the interval, that is, ∃ λ > 0 s.t.
P(N (h) = 1) = λh + o(h), as h ↓ 0.
1. W (0) = 0.
3. The increments W (t) − W (s) are Gaussian with mean zero and variance t − s,
independently of time t.
4. The sample paths t ∈ [0, ∞) → W (t, ω) of the process {W (t)} are continuous
functions of time.
Note: The white noise process ξ(t) is a stochastic process and is the derivative
of Brownian motion (that is, dW (t) = ξ(t)dt). White noise is introduced to model
or sample path of the process on the time interval [0, T ]. (Kijima, 2003)
Figure 1.1, which was simulated using the MATLAB program simbrownian.m in
Appendix E(1), shows such a path.
3. has infinite variation on any interval, no matter how small the interval is, and
Brownian motion has several applications; for quite some time, Brownian motion
has been used to model prices of risky assets such as share prices, foreign exchange
rates, and many others.
is called a Brownian motion with drift (or arithmetic Brownian motion). (Di-
ener, 2007)
11
1.5
0.5
W(t)
-0.5
-1
-1.5
0 0.2 0.4 0.6 0.8 1
b=0.2
Brownian motion (and later Brownian motion with drift) was proposed by Bachelier
in 1900 as a model for stock prices. However, this model had one drawback, namely,
that the stock prices could assume negative values (as both Figure 1.1 and Figure 1.2
clearly show). For this reason, mathematicians and financial analysts have since
resorted to the geometric Brownian motion as a better model for stock prices. This
model was first proposed by Samuelson (1965) who argued that real stock prices
could not be negative because of the limited liability of shareholders.
drift) Consider a real-valued Brownian motion with drift, S(t) = µt+σW (t), defined
on a filtered probability space. Then the sample paths of S(t) possess the following
properties:
2. Sample paths are of infinite variation on any finite interval [0, t].
3. The quadratic variations of Brownian motions with drift are finite on any in-
terval [0, t] and a.s. converge to σ 2 t in the limit n → ∞ (as the partition
(Matsuda, 2005)
Figure 1.2, simulated using the MATLAB program brownian.m in Appendix E(2),
shows a sample path of Brownian motion with drift.
20
BM with drift
quadratic variation
15
10
X(t)
-5
0 0.2 0.4 0.6 0.8 1
Drift = 10, diffusion coefficient = 5
where W (t) is a standard Brownian motion, µ the percentage drift or expected change
in S(t) and σ the percentage volatility or the variance of the change, whose solution
n 2
o
is S(t) = S(0) exp µ − σ2 t + σW (t) (Hubbard and Saglam, 2006).
tion. The process {S(t)} is a log-normally distributed random variable with mean
2
E[S(t)] = S(0)eµt and variance var[S(t)] = S(0)2 e2µt eσ t − 1 . Geometric Brow-
nian motion has applications in such fields as insurance, economics, mathematical
epidemiology, and so on, and has gained a lot of popularity in the field of finance as a
model for prices of risky assets owing to its very important property of not taking on
negative values. A sample path of geometric Brownian motion was simulated using
60
S(t)
40
20
0
0 1 2 3 4 5
= 1, = 1
N (t)
X
is called a compound Poisson process if it can be represented by S(t) = Si ,
i=1
t ≥ 0, where {N (t)}t∈R+ is a (homogeneous) Poisson process and S1 , S2 , . . . are iid
random variables that are also independent of {N (t)}t∈R+ (Boxma and Yechiali, 1995;
Kijima, 2003).
Depending on the choice of the counting process {N (t)}, there are different models
for the total claim amount process {S(t)}. For example, in the Cramer-Lundberg
model, where {N (t)} is a homogeneous Poisson process, {S(t)} is modelled as a
compound Poisson process. Another prominent model for S(t) is called the renewal
or Sparre-Andersen model, where {N (t)} is a renewal process (see Geiss (2010) for
details).
Note: The information about the asymptotic growth of the total claim amount
enables one to give advice as to how much premium should be charged in a given
time period in order to avoid bankruptcy (ruin) in the portfolio. Common classical
premium calculation principles include the net or equivalence principle, the expected
value principle, the variance principle and the standard deviation principle, all of
which are well described in Mikosch (2004).
Definition 1.1.27. (Total claim amount process) The total (or aggregate) claim
N (t)
X
amount process {S(t)} is defined as S(t) = Si , t ≥ 0, where {N (t)} is the claim
i=1
number process.
The distribution of the total claim amount process can be calculated using an exact
numerical procedure known as the Panjer recursion scheme (Mikosch, 2004). The
distribution of the claim size can also be approximated using the central limit theorem
or by means of Monte Carlo simulation techniques. Suffice it to say that the total
claim amount process {S(t)} is a compound Poisson process (see Definition 1.1.26).
Collective risk theory disregards individual claims in favour of the total gain or loss
15
of the insurance company. Therefore, this section will discuss the issues that are
addressed by collective risk theory. A claim size distribution is a distribution that
represents the sizes of claims that the insurance company has to pay. The question
that arises here is, ‘What are the realistic claim size distributions?’ In the literature,
many different types of distributions have been suggested. These are grouped into
two:
1. Light-tailed distributions
F̄ (x)
lim sup <∞
x→∞ e−λx
for some λ > 0, where F̄ (x) = 1 − F (x), x > 0, then F is referred to as light-
tailed. According to Mikosch (2004) (see Table 3.2.17), examples of light-tailed
2. Heavy-tailed distributions
These are distributions of large claims. Such claim size distributions typically
occur in a reinsurance portfolio, where the largest claims are insured. If
F̄ (x)
lim inf >0
x→∞ e−λx
This study considers a risk model comprising two distinct processes: a surplus-
generating process and an investment-generating process. The surplus-generating
process to be considered here is a diffusion-perturbated classical risk process that
determine the value of the retention percentage b that minimizes the probability of
ultimate ruin in the presence of investments and to compare this with the case when
there is no reinsurance.
is always faced with the risk of ruin. Of the many risk-reduction strategies available
to the company, it is assumed that the company uses reinsurance and investment.
In particular, it is assumed that the company takes out a quota-share proportional
reinsurance contract and invests its surplus in the money market as well as the stock
market. With these assumptions, the study seeks to find an optimal reinsurance
percentage b ∈ (0, 1], that is, the value of b that minimizes the probability of ultimate
ruin ψ b (y) or, equivalently, maximizes the ultimate survival probability φb (y) =
1 − ψ b (y).
17
The general objective of the study is to minimize the probability of ultimate ruin by
proportional reinsurance and investments.
companies.
3. There exists an optimal reinsurance strategy b ∈ (0, 1], that is, a value of the
1.6 Methodology
• The risk model is formulated theoretically with all the parameters (e.g., volatil-
ity constant, rate of premium flow, rate of return from the investments, and so
on) assumed to be unknown but from a certain set. All the theorems necessary
for the existence of solutions, as well as verification theorems, are given and
proved.
• The model is validated using parameters that are commonly used in the insur-
ance literature. Model validation has been done with the help of FORTRAN
and MATLAB.
19
1.7 Conclusion
This chapter has outlined the background, significance and objectives of the study,
as well as the problem statement, research hypotheses and format of the study. The
objectives of the study were presented as to determine the role of investments in
minimizing the probability of ultimate ruin of an insurance company, to assess the
reinsurance.
CHAPTER TWO
LITERATURE REVIEW
2.1 Introduction
The aim of this chapter is to review relevant literature on optimal control involving
proportional reinsurance and investments. In this regard, the publications reviewed
are organized into three categories: those dealing with reinsurance, those dealing
with investments and those dealing with both reinsurance and investments. These
three categories of literature are presently reviewed.
2.2 Reinsurance
According to Mikosch (2004), reinsurance treaties are of two types: Random walk
type reinsurance which includes proportional, excess-of-loss (XL) and stop-loss rein-
surance, and extreme value type reinsurance which includes largest claims and ECO-
MOR reinsurance (Excédent du coût moyen relatif or ‘excess of the average cost’).
Proportional reinsurance is a common form of reinsurance for claims of ‘moderate’
size, and requires the reinsurer to cover a fraction of each claim equal to the fraction
of total premiums the reinsurer receives from the cedent. Excess-of-loss (XL) rein-
surance involves the reinsurer paying for all individual losses in excess of some limit
time t = 0 the reinsurer guarantees to cover the k largest claims occurring in the
time period [0, t]. An ECOMOR reinsurance treaty is really an XL reinsurance with
20
21
a random deductible determined by the k-th largest claim in the portfolio (Mikosch,
2004). This study considers proportional reinsurance, which is also of two types:
quota-share and surplus reinsurance. With quota-share reinsurance the insurer and
reinsurer agree to share claims and premiums in the same proportion and this pro-
portion remains constant throughout the portfolio. In a surplus treaty the reinsurer
agrees to accept an individual risk with sum insured in excess of the direct reten-
tion limit set by the ceding company (expressed in monetary units) (Goovaerts and
Vyncke, 2004).
method to find optimal retention levels. de Finetti’s approach starts from considering
that any reinsurance policy reduces the insurer’s risk (in terms of the variance of the
random profit and the related ruin probability) as well as the expected profit. He
then proposed a two-step method: The first step involves minimizing the variance
under the constraint of a given expected profit, whereas the second one, assuming the
expected profit as a parameter, leads to the choice, based on a preference system,
is the gross premium (before expenses and reinsurance) for risk Si , and µ depends on
the cedent’s expected profit. He then concluded that if a risk is actually reinsured,
the retention is directly proportional to the safety loading and inversely proportional
of the retention for quota-share (r) and for excess-of-loss (M ). Under the usual
assumptions, he proved that r < 1 when the premium is calculated using the variance
22
principle or the exponential principle. He further proved that if the aggregate claims
are compound Poisson and the reinsurance premium is calculated according to the
expected value principle (with loading coefficient β), the optimal retention is attained
at the unique point M satisfying M = R−1 ln(1 + β), where R is the adjustment
coefficient.
In a study that complements the work of Waters (1983), Hald and Schmidli (2004)
considered the problem of maximizing the adjustment coefficient under proportional
reinsurance both for the Cramér-Lundberg model and the Sparre-Andersen model
Glineur and Walhin (2006) revisited de Finetti’s retention problem for proportional
reinsurance by applying the convex optimization method. They extended the result
to variable quota-share and surplus reinsurance with table of lines and found, by
means of a numerical example, that neither variable quota share reinsurance nor
surplus reinsurance with table of lines may be considered as optimal reinsurance
structures. They were able to determine the optimal quota-share and surplus rein-
surance strategies. However, the numerical example also led them to the conclusion
that there exists no general rule asserting superiority of either quota-share-type or
These results differ from those arrived at in an earlier study by Lampaert and Walhin
and Walhin (2005) also observed that surplus reinsurance with a table of lines based
on the inverse frequency or inverse rate method, is not optimal when compared to
surplus reinsurance with one single line. This goes against the traditional belief of
practitioners. They did not prove that this is always true but simply that a table
of lines is not always optimal. They also derived, using de Finetti’s criterion, the
optimal table of lines which their numerical example showed to be more efficient
2.3 Investments
In their study on stochastic control theory for optimal investment, Castillo and
Parrocha (2003) sought to determine the investment strategy that minimizes infinite
time ruin probability. In this regard, they considered an insurance business with a
fixed amount available for investment in a portfolio consisting of one risk-free asset
and one risky asset and gave a numerical algorithm for solving the resulting HJB
equation. Assuming exponential claim-sizes, they used several numerical examples to
derive the optimal investment portfolio for different investment scenarios. Different
from Liu and Yang (2004) the examples showed that, assuming an amount A(t)
available for investment, the fraction of A(t) invested in a risky asset is directly
proportional to the insurer’s surplus. They also concluded, based on the numerical
results, that the optimal combinations of the risk-free and risky asset varies with the
value of A(t) allotted for investment, and that no single optimal combination can
Liu and Yang (2004) studied the problem of optimal investment for an insurer to
24
minimize its probability of ruin. Their model was a generalization of the model in
Hipp and Plum (2000) and assumed that the insurance company receives premiums
at a constant rate and that it can invest in the money market and in a risky asset
such as stocks. They investigated the investment behaviour numerically for various
claim-size distributions and computed the optimal investment policy and the solution
of the associated HJB equation under each assumed claim-size distribution. They
also investigated the effects of changes in various factors, such as stock volatility, on
optimal investment strategies and survival probability. Furthermore, they extended
their model to allow for cases in which borrowing constraints or (stop-loss) rein-
surance are present. They found that with exponentially distributed claim-sizes the
insurer’s surplus is inversely proportional to the level of investment in the stock mar-
ket, whereas with a Pareto claim-size distribution the optimal investment strategy
is to invest more in stocks when the initial surplus is large. As for the effects of un-
derlying factors on the survival probability, they found that the survival probability
is directly proportional to the risk-free interest rate, the expected rate of return of
stocks, the safety loading and the expected claim-size, but inversely proportional to
the stock price volatility and the intensity λ of the claim number process N (t). Their
study did not explore the possibility of proportional reinsurance and what impact
it would have on the probability of survival (and, therefore, of ruin) in the light of
those various underlying factors.
Paulsen et al. (2005) considered the problem of minimizing the infinite time ruin
probabilities for the diffusion-perturbated classical risk process compounded by a
linear Brownian motion and allowed for stochastic return on investments. They gave
sufficient conditions for the survival probability function to be four times continu-
ously differentiable, which in particular implies that the survival probability is the
conjunction with Simpson’s rule. This study, though allowing for investment, does
not incorporate reinsurance of any kind.
Schmidli (2002) studied the problem of minimizing the probability of ruin by invest-
ment and reinsurance. He considered a classical risk model and allowed investment
into a risky asset modelled as a Black-Scholes model as well as proportional reinsur-
ance. Using the HJB approach, he found optimal levels of investment and reinsurance
which minimize the ruin probability. Among other things, he found that, for Pareto
distributed claim sizes (for the case where the whole insurance risk is reinsured), in-
vestment and reinsurance decrease the ruin probability considerably for larger initial
capital. But, again, this study assumed investment in a risky asset only.
Irgens and Paulsen (2004) studied the problem of maximizing the expected utility
of the assets of an insurance company at a terminal time by factoring reinsurance
the possibility that the company could cover some of its risk using a proportional
reinsurance contract and the remaining risk using XL reinsurance. They used three
different utility functions: power utility with termination at ruin, logarithmic utility
and exponential utility. By means of the Hamilton-Jacobi-Bellman (HJB) equation
they verified that the suggested solutions were in fact optimal among a fairly large
easily verifiable admissible class of controls. However, in addition to using two types
of reinsurance (proportional and XL), their study did not include the fact that the
insurance company could invest in both a risk-free and a risky asset.
Ma et al. (2008) considered the problem of minimizing the probability of ruin under
26
interest force. Their study focused on the classical risk process and allowed for
the possibility of investing in a risk-free asset as well as purchasing proportional
reinsurance. They found that the optimal proportional reinsurance did actually
minimize the probability of ultimate ruin (by maximizing the survival probability).
They derived the HJB equation but found, differently from Schmidli (2001) and
Schmidli (2002), that it did not always have a smooth solution. For this reason,
they considered two cases. The first case was a trivial case in which they found the
corresponding minimal probability of ultimate ruin and the optimal proportional
reinsurance strategy directly. In the second case, they used the HJB equation to
obtain the minimal probability of ruin and the optimal proportional reinsurance
strategy by means of a new verification theorem. It should be noted that the study
by Ma et al. (2008) did not include the possibility of investing in a risky asset as
well.
In their study, Kim and Lee (2008) sought to determine stochastic optimal rein-
surance and investment strategies for the surplus of an insurance company. They
applied proportional reinsurance and investments to the Cramér-Lundberg model
u(t) ∈ [0, 1] is the risk exposure or retention level. This model is the same as the one
in Højgaard and Taksar (1998) who considered a proportional reinsurance policy π
applied to the surplus processRπ (t) of an insurance company governed by an SDE
of the insurer’s surplus in a financial market with two stocks whose prices were de-
scribed by linear SDEs. The purpose of their paper was to give an explicit expression
for the optimal reinsurance and investment strategy which maximizes the expected
27
exponential utility of the final value of the surplus of a life insurance company at the
end of the T -th year. To do this, they derived and solved the corresponding HJB
equation. This study only considered investment of the surplus in risky assets and
did not include the case where part of the surplus could be invested in a risk-free
asset.
Paulsen’s (2008) study on ruin models with investment income is a survey that
treats the problem of ruin in a risk model when assets earn investment income. He
introduces the risk process by means of a basic insurance process and an investment-
bounds on the ruin probability, finally exploring the possibility of minimizing the
ruin probability by investment and possibly reinsurance control. In particular, he
considers the case where the insurer has the possibility of purchasing a proportional
as those in Paulsen and Gjessing (1997b) and Paulsen et al. (2005), forms the basis
of the model to be studied in the research currently being undertaken.
Liu and Ma (2009) studied optimal reinsurance and investment problems for gen-
eral insurance models. Their study considered the utility optimization problem for
an insurance company whose surplus process is described by an SDE driven by a
Brownian motion and a Poisson random measure. Their optimization problem was
based on the consideration that the company can manage its surplus, hence risk,
by investments, proportional reinsurance and consumption. It is assumed that the
company invests its surplus in one risk-free asset and in some risky assets and that
it is allowed to change its investment positions continuously. Additionally, Liu and
28
Ma (2008) assumed that the insurance company can cede a fraction of the incoming
claims, while at the same time yielding a fraction of its premium, to a reinsurance
company. Finally, the insurance company is also allowed to ‘consume’ (in the form of
dividend, refund, etc.). Rather than use the HJB equation approach and optimizing
the survival probability, they used the utility optimization approach by optimizing
a ‘truncated’ utility function with the help of the theory of backward stochastic
differential equations (BSDEs).
and Thonhauser (2009) discussed several model extensions and possibilities of con-
trol actions on the insurance portfolio surplus process which included reinsurance
and investment. They discussed investment in a risky asset and underscored the role
of proportional and XL reinsurance as control variables for minimizing the ruin prob-
ability but because their interest was in dividends, they did not discuss investment
and reinsurance to any detail.
which requires appropriate regularity of the value function as well as other properties
that are not always easy to verify and for which analytic solutions are difficult to
come by, their study focused on value iteration that does not require these regularity
assumptions and that, in certain cases, allows for an analytic solution. In those
cases where an analytic solution may be difficult to obtain, value iteration makes
it possible to compute the optimal values and control to any degree of accuracy.
In fact, the optimal value turns out to be the fixed point of a suitable contraction
operator. The solution therefore involves the determination of this fixed point that
can be arbitrarily closely approximated by iterating this operator sufficiently often.
Thus Edoli and Runggaldier (2010) showed that, by choosing an exponential utility
function that combines features of ruin minimization and utility maximization, it
29
is possible to obtain a semianalytic solution to the problem in the sense that the
solution to the fixed point problem can be expressed in terms of two Volterra integral
equations that can be solved explicitly. They then obtained the optimal investment
and proportional reinsurance strategies numerically.
None of the literature reviewed in this chapter deals with the retention percentage b ∈
(0, 1] for reinsurance as the only control variable in the presence of investments (as the
current study proposed to do). Those dealing with reinsurance and investments all
consider both the reinsurance and investment strategies as control variables. In other
words, they determine the values of the retention level b ∈ (0, 1] and the investment
level a ∈ [0, 1] (where a is the proportion of the insurer’s surplus invested in a
risky asset) that minimize the probability of ultimate ruin of an insurance company.
Furthermore, to the researcher’s knowledge, no study exists prior to this that applies
the block-by-block method to optimal control problems involving reinsurance. Hence
the importance of this study.
2.5 Conclusion
This chapter of the study has reviewed the relevant literature on proportional rein-
surance and investments. The next chapter outlines the model formulation process
and states the model to be studied.
CHAPTER THREE
MODEL FORMULATION
3.1 Introduction
In this chapter, the model for the study is formulated and the model formulation
process is clearly outlined. This study considers a jump-diffusion model with rein-
surance and investments. But before looking at this model it is necessary to state
The theoretical foundations of modern risk theory were laid in 1903 by the Swedish
actuary Filip Lundberg. One of Lundberg’s major contributions was the introduction
that the insurer’s surplus level ever falls below zero, thus making the firm technically
bankrupt. Lundberg’s model was subjected to rigorous mathematical treatment
about 30 years later by the famous Swedish actuary and probabilist Harald Cramér
(1930). For this reason, this model has come to be called the Cramér-Lundberg
model (or the classical risk process). Cramér extensively developed collective risk
theory, also called ruin theory, by using the total claim amount process S(t) with
30
31
where
• p is the premium rate (that is, the the insurer’s premium income per unit time
assumed to be received continuously) which may be given by p = (1 + η)λµ,
where η is the insurer’s safety loading.
F (s). This study considers Exp(µ) for small claims so that F (s) = 1 − e−µs
κ α
and Pareto(α, κ) for large claims so that F (s) = 1 − κ+s .
PN (t)
• i=1 Si (usually denoted by S(t)) is a compound Poisson process with an
average number of claims per time period of λ. This process (also called the
total claim amount process) represents the aggregate losses to the insurance
company.
The classical risk process in (3.1) is the surplus process without reinsurance and has
dynamics given by
N (t)
X
dY (t) = pdt − d Si (3.2)
i=1
32
with initial reserve or surplus Y (0) = y. The Cramér-Lundberg model with reinsur-
ance will be discussed as one of the cases of the jump-diffusion model when certain
where p is the premium rate, that is, the insurer’s premium income per unit time
assumed to be received continuously. The first term therefore represents the ex-
pected premium income of the insurance company. WP is a standard Wiener process
independent of the compound Poisson process N
P P (t)
i=1 SP,i , {SP,i } is the aggregate
claims. The claim sizes are a sequence of strictly positive independent and identi-
cally distributed (iid) random variables {SP,i }i∈N which are independent of the claim
arrival and claim number processes. We denote by FP the distribution function of
SP,i , by µP = E[SP,i ] its mean value and by MS (r) = E erSP,i its moment-generating
function. We will assume that FP (0)=0 and that at least one of σP or λP is non-zero.
the classical risk process (or the Cramér-Lundberg model). If the initial surplus of
the insurance company is y, then equation (3.3) is referred to as the classical risk
process perturbed by a diffusion or a diffusion-perturbated classical risk process (e.g.
in Furrer and Schmidli, 1994; Irgens and Paulsen, 2004; Cai and Xu, 2006).
33
But equation (3.3) represents the insurer’s surplus process without reinsurance. Now
suppose that the insurer can take out reinsurance. In other words, suppose it is able
to cede some of the risk it carries to a reinsurer. Then the company can only do so
at the expense of a share of the premiums it receives from its clients. Reinsurance
was defined in definition 1.1.2 and is of two types: Random walk type reinsur-
ance which includes proportional, excess-of-loss (XL) and stop-loss reinsurance, and
extreme value type reinsurance which includes largest claims and ECOMOR rein-
surance (Excédent du coût moyen relatif or ‘excess of the average cost’). A full
description of these types of reinsurance is given in Mikosch(2004), Geiss (2010),
Goovaerts and Vyncke (2004) and Centeno and Simões (2009) (see also Chapter
reinsurance which, as Lampaert and Walhin (2005) point out, is the easiest way of
covering an insurance portfolio.
With proportional reinsurance, the first-line insurer (which is the ceding company
or cedent) and the reinsurer agree on a cession percentage for each policy in the
insurance portfolio. These two parties may also have to agree as to which type
proportional reinsurance above the other. For this reason, this study restricts itself
to quota-share proportional reinsurance.
The cession percentage represents how both premiums and claims will be shared
between the two parties. Let the stochastic process {b(t)}t∈R+ , where b(t) is the
34
retention level at time t for reinsurance, represent a reinsurance strategy with values
in (0, 1]. This means that the insurer pays b(t)SP,i of the i-th claim SP,i occurring
level, as well as the cession percentage, is constant over the entire insurance portfolio,
that is, b(t) = b for all risks occurring at any time t (Paulsen, 2008; Schmidli, 2002).
Thus under a quota-share proportional reinsurance contract, the reinsurer is liable
for (1 − b)SP,i while the insurer covers bSP,i of the i-th claim occurring at time t.
The factor (1 − b) is the cession percentage for the reinsurance which represents
the proportion of claims ceded to the reinsurer as well as the share of premiums the
reinsurer is to receive from the first-line insurer (Glineur and Walhin, 2006; Lampaert
and Walhin, 2005; Schmidli, 2001; Ma et al., 2008).
becomes
NP (t)
X
b
P (t) = bpt + bσP WP (t) − bSP,i (3.4)
i=1
having dynamics
NP (t)
X
dP b (t) = bpdt + bσP ξ(t)dt − d bSP,i (3.5)
i=1
where ξ(t) is a white noise process. Assuming an initial surplus y and setting σP
to zero in equation (3.4) yields the Cramér-Lundberg model with reinsurance. This
(the classical risk process with reinsurance) is the model which was studied by Ma
et al. (2008) who obtained the minimal probability of ruin as well as the optimal
35
study.
The classical risk model represented by equation (3.1), though simple, is rather
inadequate for modelling real-world insurance processes. As Taylor and Buchanan
(1988) have observed, ‘While this theory is well developed and well known, there
are a number of respects in which it lacks realism to a point which militates against
its practical use without substantial modification.’ One of the major limitations of
the Cramér-Lundberg model is that it does not account for interest earned on the
reserve (Hipp, 2003; Kasozi and Paulsen, 2005b). In other words, it assumes that the
insurance company does not earn any interest on its surplus. In addition, according
to Hipp (2004), the Cramér-Lundberg model does not account for long tail business
with claims that are settled long after occurrence of the claim, nor does it include
time-dependence or randomness of premium income and of the size of the portfolio.
This study deviates from the classical setting and assumes that in addition to rein-
surance the company invests some of its surplus in a risk-free asset such as a bond at
a positive risk-free rate r0 and the rest in a risky asset such as a stock or market in-
dex described by the geometric Brownian motion. In other words, for simplicity, we
consider only two assets in the financial market: a risk-free bond and a risky asset,
namely stocks (Liu and Yang, 2004). The first attempt to incorporate investments
tion was that capital earned interest at a fixed rate. This assumption also appears in
Paulsen (2003). Other researchers have expanded on this area since then and it has
increasingly become popular (for example, Black and Scholes (1973) whose model
is discussed below). Paulsen and Gjessing (1997b) considered a risk process with
stochastic return on investments. This study seeks to modify Paulsen and Gjessing’s
dynamics is:
dB(t) = r0 B(t)dt
where B(t) is the price of the risk-free bond at time t, and r0 > 0 is the risk-free
interest rate.
The price of the stock is the process S(t) which is modelled by a geometric Brownian
motion with dynamics:
where S(t) is the price of the stock at time t, r is the expected instantaneous rate of re-
turn of the stock (r > 0), σR is the volatility of the stock price (σR > 0), {WR (t)}t∈R+
is a standard Brownian motion defined on the probability space (Ω, F, P) and inde-
pendent of {WP (t)}t∈R+ (Liu and Yang, 2004), and dWR (t) = ξ(t)dt (ξ(t) being
white noise). The risky asset (stock) yields a higher return than the risk-free bond,
that is, r0 < r.
37
Combining all these into a return on investments process (as in Paulsen and Gjessing,
1997b) gives the following:
NR (t)
X
R(t) = rt + σR WR (t) + SR,i , t ≥ 0, R(0) = 0 (3.6)
i=1
where WR (t) is another Brownian motion independent of the surplus process P (t)
and N
P R (t)
i=1 SR,i is another compound Poisson process. Taking λR > 0 makes the
model more realistic in that it ensures that sudden changes in the value of the stock
are taken into account; however, it will be assumed here that λR = 0, so that we
have the Black-Scholes model:
In other words, the insurance company can invest part of its surplus in a risk-free
asset and the rest in a risky asset. The interpretation of (3.7) is that r is the risk-
free part of the investments so that R(t) = rt implies that one unit invested at time
zero will be worth ert at time t. The term σR WR (t) then takes account of random
ments
The risk process after taking out reinsurance and obtaining investment returns is a
combination of the two processes in equations (3.4) and (3.7). It is the process Y b =
{Y b (t)}t∈R+ which represents the insurance portfolio, that is, the insurance process
compounded by the return on investment process and by proportional reinsurance.
Versions of this process have been extensively studied for ultimate ruin probability
by several authors (see, for example, Paulsen et al. (2005), Paulsen (1998), Paulsen
(2008), Paulsen and Gjessing (1997b), Paulsen and Rasmussen (2003) and Sundt
38
and Teugels (1995)). That is, the risk process has the dynamics
where the non-negative constant y = Y b (0) > 0 is the initial capital or surplus of
the insurance company, P b (t) is the basic insurance (or surplus-generating) process
in equation (3.4), R(t) the investment process in equation (3.7) and Y b (t− ) denotes
the insurer’s surplus just prior to time t.
In summary, this work analyses the model represented by equations (3.4), (3.7) and
(3.9):
NP (t)
X
P b (t) = bP (t) = bpt + bσP WP (t) − bSP,i
i=1
(3.10)
R(t) = rt + σR WR (t)
Z t
b b
Y (t) = y + P (t) + Y b (s− )dR(s)
0
3.7 Conclusion
In this chapter the model to be studied has been formulated. This is equation (3.9)
which comprises the surplus process (equation (3.4)) and the investment generating
process (equation (3.7)). The next chapter analyses the model by way of stating the
4.1 Introduction
In this chapter the model ((3.10) comprising equations (3.4), (3.7) and (3.9)) is
analyzed. The problem’s HJB equation is stated, as is the corresponding Volterra
integro-differential equation (VIDE) which is then transformed into a linear Volterra
Definition 4.1.1. (Time of ruin) This is the first time that an insurer’s surplus
process becomes negative and is mathematically defined as
τb = inf t > 0 : Y b (t) < 0|Y b (0) = y ,
with τb = ∞ if Y b (t) remains positive, where Y b (t) is the risk process (equation (3.9)
above) incorporating quota-share proportional reinsurance with strategy b(t) = b ∈
Definition 4.1.2. (Probability of ultimate ruin) This is the probability that the
surplus of an insurance company ever drops below zero (Burnecki and Miśta, 2005)
ψ b (y) = P τb < ∞|Y b (0) = y = 1 − φb (y)
39
40
‘What value of the retention level b should the insurer choose in order to minimize
the probability of ruin, or in order to maximize the probability of survival?’ This is
the question to be answered in this work. The study seeks to maximize the survival
∗
and, hopefully, the optimal reinsurance strategy {b∗ (t)} s.t. φ(y) = φb (y). The
quantity φ(y) in equation (4.1) is the value function associated with the reinsurance
strategy b ∈ (0, 1]. The survival probability is therefore the objective function and
the control variable to be adjusted so that the objective function is maximized is
the rentention percentage b ∈ (0, 1] for the quota-share proportional reinsurance.
ruin).
where
r − 21 σR2 t + σR WR (t) ,
R̄(t) = exp t≥0
Since the investment generating process R(t) follows (3.7), it follows that under weak
(where F̄ (y) = 1 − F (y)) with boundary conditions lim ψ(y) = 0 and ψ(y) = 0 if
y→∞
Sometimes it is more convenient (as we do in this study) to work with the survival
Aφ(y) = 0.
(Y b solves the linear SDE in (3.9)). It is therefore necessary to resort to the following
simple but useful result which was proved in Paulsen and Gjessing (1997b) and
Constantinescu (2003).
Theorem 4.1.1. Let τb = inf t > 0 : Y b (t) < 0|Y b (0) = y be the time of ruin
ψ(y) = 1 on y < 0,
lim ψ(y) = 0,
y→∞
42
then
Note: Paulsen et al. (2005) used the probability of survival φ(y) = 1 − ψ(y)
with boundary conditions
φ(y) = 0 on y < 0,
lim φ(y) = 1
y→∞
hand derivative. If qα (y) solves Aqα (y) = αqα (y) on y > 0, together with the
boundary conditions
qα (y) = 1 on y < 0,
lim qα (y) = 0,
y→∞
then
qα (y) = E[eατb ]
The HJB equation is a partial differential equation which is central to optimal control
theory. The solution to the HJB equation is the value function that needs to be
optimized. In this study, the survival probability φ(y) = 1 − ψ(y) is the value
43
thus minimizing the ultimate ruin probability (subject to the boundary conditions
in Part 1 of Theorem 4.1.1). The HJB equation for the problem is motivated as
follows (Schmidli, 2008):
Let (0, h] be a small interval, and suppose that for each surplus y(h) > 0 at time h we
ε
have a reinsurance strategy bε s.t. φb (y(h)) > φ(y(h)) − ε. We let b(t) = b ∈ (0, 1]
for t ≤ h. Then
ε
φ(y) ≥ φb (y) = E φb Y b (h) 1{τb >h}
ε
= E φb Y b (τb ∧ h)
≥ E φ Y b (τb ∧ h) − ε
(4.8)
where a(s, b) = bs denotes the part of the claim SP,i paid by the cedent. Substituting
(4.9) into the expected value (4.8) yields
Z τb ∧h
1 2 2
(ry + bp)φ0 Y b (s) + σR y + b2 σP2 φ00 Y b (s)
E
0 2
Z y
b b
+λ φ Y (s) − bs dF (s) − φ Y (s) ds ≤ 0
0
Dividing by h and letting h → 0 yields, provided that limit and expectation are
interchangeable,
Z y
0 1 2 2
σR y + b2 σP2 φ00 (y) + λ
(ry + bp)φ (y) + φ (y − bs) dF (s) − φ(y) ≤ 0
2 0
44
≤0 (4.10)
Suppose that there is an optimal strategy b ∈ (0, 1] s.t. lim b(t) = b(0). Then, as
t↓0
above,
Z τb ∧h
1 2 2
(ry + bp)φ0 Y b (s) + σR y + b2 σP2 φ00 Y b (s)
E
0 2
Z y
b b
+λ φ Y (s) − bs dF (s) − φ Y (s) ds = 0
0
1 2 2
Dividing by h and letting h → 0 yields (ry + b0 p)φ0 (y) + σR y + b20 σP2 φ00 (y) +
Z y 2
λ φ(y − b0 s)dF (s) − φ(y) = 0
0
=0 (4.11)
with boundary conditions φ(y) = 0 on y < 0 and lim φ(y) = 1 (see Theorem 4.1.1).
y→∞
The function φ(y) will satisfy (4.11) only if φ(y) is strictly increasing, strictly concave,
twice continuously differentiable and satisfies φ(y) → 1 for y → ∞ (Hipp and Plum,
2000). In the following, therefore, φ(y) will be assumed to be strictly increasing.
This is consistent with the smoothness assumption and the intuition that the more
wealth there is (through investment), the higher the probability that the insurer
will survive. It will also be assumed that φ(y) is concave. To ensure smoothness
and concavity, the claim density function must be locally-bounded (Liu and Yang,
2004). The following results provide verification of the existence and property of the
solution of the HJB equation (4.11).
The proof of this result has been given in Hipp and Plum (2003).
strictly concave.
This result has been proved in Hipp and Plum (2000) and Schmidli (2002).
The integro-differential equation for the survival probability φ(y), which follows from
the HJB equation (4.11), is of the form Aφ(y) = 0 (since, by (4.6), φ(y) = 0 for
y < 0), where A is the infinitesimal generator (4.4) of the underlying risk process
for 0 < y ≤ ∞. Henceforth we drop the subscript in λP and simply write λ. Equation
(4.12) could then be rewritten as
Z y
1 2 2
σR y + b2 σP2 φ00 (y) + (ry + bp)φ0 (y) + λ
φ(y − bs)dF (s) − λφ(y) = 0, (4.13)
2 0
where, for the case with no diffusion (i.e., when σP2 = σR2 = 0) and with no reinsur-
ance (i.e., when b = 1),
r + λF̄ (y − s)
K(y, s) = − ,
ry + p
(4.15)
p
α(y) = φ(0),
ry + p
with F̄ (s) = 1 − F (s). When there is diffusion (i.e., when σP2 + σR2 > 0), but with no
Rs
with F2 (s) = 0
F (v)dv
But when there is reinsurance, for the case with no diffusion, we have
r + λ (1 − bF (y − bs))
K(y, s) = − ,
ry + bp
(4.17)
bp + λbF2 (s)
α(y) = φ(0),
ry + bp
Proof. The proof for the case without reinsurance is given in Paulsen et al. (2005).
Here we present the proof for the case with proportional reinsurance. Integrating
equation (4.13) by parts w.r.t. to y on [0, z] gives
Z z
1 2 2 0 1 2 2 0
2 2
σ z + b σP φ (z) − b σP φ (0) − σR 2
yφ0 (y)dy + (rz + bp))φ(z) − bpφ(0)
2 R 2
Z z Z z Z y0 Z z
−r φ(y)dy + λ φ(y − bs)dF (s)dy − λ φ(y)dy = 0
0 0 0 0
(4.18)
1 2 2 1
σR z + b2 σP2 φ0 (z) − b2 σP2 φ0 (0) + (r − σR2 )z + bp φ(z) − bpφ(0)
2 Z2 z Z zZ y
2
− r + λ − σR φ(y)dy + λ φ(y − bs)dF (s)dy = 0.
0 0 0
(4.19)
To simplify the double integral in (4.19) we again use integration by parts and
switch the order of integration using Fubini (Günther, 2008; Schmidli, 2008). Recall
that F (0) = 0 and F (s− ) = F (s) for s ∈ R (F being absolutely continuous w.r.t.
Lebesgue measure). So
Z z Z y Z z Z y
0
φ(y − bs)dF (s)dy = φ(y − by)F (y) − φ(y)F (0) + b F (s)φ (y − bs)ds dy
0 0
Z0 z Z zZ y 0
1 2 2 1
σR y + b2 σP2 φ0 (y) − b2 σP2 φ0 (0) + (r − σR2 )y + bp φ(y) − bpφ(0)
2 Z y Z2 y Z y
2
− r + λ − σR φ(s)ds − λb φ(0)F (s)ds + λb F (y − bs)φ(s)ds = 0
0 0 0
(4.21)
or
1 2 2 1
σR y + b2 σP2 φ0 (y) − b2 σP2 φ0 (0) + (r − σR2 )y + bp φ(y) − bpφ(0)
2 Z2 y
Z y (4.22)
λbF (y − bs) − r + λ − σR2 φ(s)ds = 0
−λb φ(0)F (s)ds +
0 0
48
If there is no diffusion, i.e., if σP2 = σR2 = 0, then this is simply (4.14) and (4.17). If
either σP or σR is non-zero, then integrating (4.22) by parts over y on [0, z] yields
For illustrative purposes, this study will restrict itself to reinsurance in the non-
diffusion case. The following result has been proved in Ola (2006).
Theorem 4.3.2. The Volterra integral equation of the second kind (4.14) has a
unique solution φ ∈ C ([c, d]) for any continuous α.
4.4 Conclusion
In this chapter the problem’s HJB equation has been stated as well as its correspond-
ing Volterra integro-differential equation (VIDE) which was then transformed into a
linear Volterra integral equation (VIE) of the second kind. Chapter Five proposes
a numerical method for solving this VIE and presents some numerical results based
on light- and heavy-tailed claim-size distributions.
CHAPTER FIVE
NUMERICAL RESULTS
5.1 Introduction
This chapter presents a numerical method for solving the VIE (4.14) and some nu-
merical results based on this method both for small and large claims. In this regard,
results are presented based on the exponential and Pareto distributions for small and
This section discusses numerical solutions of the survival probability φ(y) using a
fixed grid y = 0, h, 2h, . . .. It is assumed throughout that the assumptions of Theorem
4.1.1 hold. For this to happen, by Theorem 4.3.1, it is necessary that r > 12 σR2 (since,
by Paulsen (1998), if φ(y) is the survival probability and p > 0, r > 0 and λ > 0,
then φ(y) > 0 iff r > 12 σR2 , and in this case φ(∞) = 1. When r ≤ 21 σR2 , φ(y) = 0 ∀ y).
The numerical solution of the general linear Volterra integral equation of the second
kind
Z y
g(y) + K(y, s)g(s)ds = β(y), (5.1)
0
where the kernel K(y, s) and the forcing function β(y) are known functions and g(y)
is the unknown function to be determined, is of the form
n
X
gn + h wi Kn,i gi = βn (5.2)
i=1
where gi is the numerical approximation to g(ih), Kn,i = K(nh, ih), gn = g(nh) and
βn = β(nh). The wi are the integration weights. Here, the block-by-block (B-by-
49
50
known to have an error of order 4, see Theorem 5.2.2 and Remark 5.2.3) to obtain
solutions in blocks of two values.
According to Press et al. (1992), the general consensus is that the B-by-B method,
which was first proposed by Young (1954), is the best of the higher order methods for
solving (5.1). The B-by-B methods are essentially extrapolation procedures which
produce a block of values at a time and have the advantage over linear multistep
and step-by-step methods in that they can be of high order and still be self-starting
(Makroglou, 1980; Saeed and Ahmed, 2008; Delves and Mohamed, 1985; Delves and
Walsh, 1974). Apart from not requiring special starting procedures, B-by-B methods
are simple to use and switching step-size presents no problem (Linz, 1969).
Linz (1969) describes a two-block B-by-B method which he used to solve non-linear
Volterra integral equations of the second kind. AL-Asdi (2002) also used two- and
three-block methods for solving Hammersetien-Volterra integral equations of the
second kind, while Saify (2005) solved a system of linear Volterra integral equations
of the second kind using two-, three- and four-block methods. Makroglou (1980)
applied a B-by-B method to the solution of non-linear Volterra integro-differential
equations (VIDEs) with continuous kernels and later (1981) adapted it to VIDEs
with weakly singular kernels. More recently, Kasozi and Paulsen (2005a) used the
two-block B-by-B method to study the flow of dividends under a constant force
of interest. They derived a linear Volterra integral equation of the second kind
and applied an order-four B-by-B method of Paulsen et al. (2005a) in conjunction
with Simpson’s rule to solve the Volterra integral equation for the optimal dividend
barrier.
In another study, Kasozi and Paulsen (2005b) applied an order-four B-by-B method
to the numerical solution of the Volterra integral equation for ultimate ruin in the
Cramér-Lundberg model compounded by a constant force of interest. More pertinent
literature on the B-by-B method is available, for example, in Paulsen (2003) and
51
Paulsen and Gjessing (1997a). But as yet the B-by-B method has not been used in
stochastic control problems involving reinsurance, hence its adoption in the current
study.
To briefly describe the method, Simpson’s rule gives, for any k ∈ C 4 [ih, (i + 2)h],
Z (i+2)h
h
k(s)ds = (k(ih) + 4k((i + 1)h) + k((i + 2)h)) + O(h5 )
ih 3
h
g2 + (K2,0 g0 + 4K2,1 g1 + K2,2 g2 ) = β2 (5.3)
3
Here, g1 is unknown, but using the same rule with gridsize h2 , Simpson’s rule gives
h
g1 + K1,0 g0 + 4K1, 1 g 1 + K1,1 g1 = β1 (5.4)
6 2 2
yields
h 3 3 1
g1 + K1,0 g0 + 4K1, 1 g0 + g1 − g2 + K1,1 g1 = β1 ,
6 2 8 4 8
that is,
h 3 1
g1 + K1,0 + K1, 1 g0 + K1,1 + 3K1, 1 g1 − K1, 1 g2 = β1 , (5.5)
6 2 2 2 2 2
Equations (5.3) and (5.5) are a pair of equations to solve for g1 and g2 . Continuing
in this manner in blocks of 2 with wi = 31 {1, 4, 2, ..., 2, 4, 1}, i = 0, 1, . . . , 2m, we get
2m
X h
g2m+2 + h wi K2m+2,i gi + (K2m+2,2m g2m + 4K2m+2,2m+1 g2m+1 + K2m+2,2m+2 g2m+2 )
i=0
3
= β2m+1 (5.6)
and
2m
X h
g2m+1 + h wi K2m+1,i gi + K2m+1,2m g2m + 4K2m+1,2m+ 1 g2m+ 1 + K2m+1,2m+1 g2m+1
i=0
6 2 2
= β2m+1 (5.7)
52
Now, g2m+ 1 is not known, but may be estimated by quadratic interpolation. Thus,
2
approximating g2m+ 1 by 83 g2m + 34 g2m+1 − 18 g2m+2 and inserting this into (5.7) yields
2
To solve each block from (5.6) and (5.7) for g2m+1 and g2m+2 , we make use of their
form, AG = β, where the 2 x 2 matrix A has entries
h h
a11 = 1 − K2m+1,2m+ 1 − K2m+1,2m+1
2 2 6
h
a12 = K 1
12 2m+1,2m+ 2
4h
a21 = − K2m+1,2m+2
3
h
a22 = 1 − K2m+2,2m+2
3
2m
X h
β2 = g2m+2 + h wi K2m+2,i gi + K2m+2,2m g2m
i=0
3
G = A−1 β
1
(g2m+1 , g2m+2 )T = (β1 a22 − β2 a12 , β2 a11 − β1 a21 )T
d
That is,
β1 a22 − β2 a12
g2m+1 =
d
β2 a11 − β1 a21
g2m+2 =
d
iff
as h → 0, N → ∞, s.t. N h = a.
is the largest number for which there exists a finite constant C s.t.
We need to show that the method given by equations (5.6) and (5.7) converges and
also establish its order of convergence. We need the following lemma which forms
Theorem 5.2.2. The approximation method given by equations (5.6) and (5.7) is
convergent and its order of convergence is four.
Remark 5.2.3. By Theorem 3.1 in Paulsen et al. (2005) and from results in Chapter
7 of Linz (1985), it follows that for a fixed y so that nh = y, the solution satisfies
provided that g is four times continuously differentiable as is the case here by The-
orem 2.4 in Paulsen et al. (2005). On the other hand, for the B-by-B method
|g2m+2 − g2m+1 | = O(h4 ) as well.
54
This section presents some numerical results obtained using the block-by-block method
outlined in Section 5.2 above. This will be done for the exponential distribution (for
small claims) and the Pareto distribution (for large claims). The ruin probabilities
for the Cramér-Lundberg model are computed and compared with those for the case
when the C-L model is compounded by reinsurance and a constant force of interest
both for the non-diffusion and diffusion cases. For illustrative purposes, the impact
of proportional reinsurance is investigated only for the non-diffusion case, though
For a given y, ψh (y) is the calculated ruin probability when a gridsize h is used.
Exp(µ) refers to the exponential density f (s) = µe−µs . Thus the distribution function
for the exponential distribution is F (s) = 1 − e−µs and F̄ (s) = 1 − F (s) = e−µs . The
1
mean excess function for the Exp(µ) distribution is eF (s) = µ
, so that F2 (s) =
s − µ1 F (s). Since in the case of the exponential distribution the exact solution ψ(y)
is known for the C-L model, it is possible to compute the absolute percentage relative
error as
ψh (y) − ψ(y)
Dh (y) = 100
ψ(y)
where ψ(y) for Exp(µ) distributed claims is given in Rolski et al. (1998) and Hipp
(2004) by the following equations:
1 −Ry 1
ψ(y) = e ; ψ(0) = (5.10)
1+Λ 1+Λ
pµ Λ
where Λ = λ
− 1 and R = µ 1+Λ . The parameter R is referred to as the adjustment
coefficient or Lundberg’s constant. Equation (5.10) shows that ψ(0) depends only
on the expected claim size and not on the specific form of the claim size distribu-
tion F (s). This distribution makes it possible to gauge the accuracy of the B-by-B
method. Hence the exponential distribution is used extensively in the examples that
follow. However, the Pareto distribution will also be used for modelling large claims.
55
ακα
f (s) = , (5.11)
(κ + s)α+1
κ α
where α > 0, κ = α − 1 > 0 and the distribution function F (s) = 1 − κ+s . Hence
κ α κ κ
the tail distribution is F̄ (s) = κ+s . Also, F2 (s) = s − 1 + κ+s . Note also that
κ+s
the Pareto distribution has a mean excess function eF (s) = α−1 (or 1 + κs ), meaning
that F2 (s) can alternatively be written as s − 1 + κs F (s). All the calculations in
this section were performed on a Toshiba Satellite L300 PC with an Intel Celeron
550 processor at 2.0GHz and 1.0GB internal memory. The B-by-B method was
implemented using the FORTRAN programming language and taking advantage of
its Double Precision feature to obtain satisfactory accuracy. Slower programs like
Splus, R, MATLAB, Maple or Mathematica could, of course, have been used but at
the expense of considerably longer computing time.
Integrating by parts on [0, y] transforms (5.12) into a VIE of the second kind:
Z y Z y
p (φ(y) − φ(0)) + λ F (y − s)φ(s)ds − λ φ(s)ds = 0
0
Z y Z 0y
pφ(y) = pφ(0) + λ φ(s)ds − λ F (y − s)φ(s)ds
0 0
Z y
= pφ(0) + λ (1 − F (y − s)) φ(s)ds
0
Z y
= pφ(0) + λ F̄ (y − s)φ(s)ds
0
λ y
Z
φ(y) = φ(0) + F̄ (y − s)φ(s)ds
p 0
that is, (5.12) can be expressed as an ordinary VIE of the second kind
Z y
φ(y) = g(y) + K(y, s)φ(s)ds
0
λF̄ (y−s)
where the forcing function g(y) = φ(0) and the kernel K(y, s) = p
(with
F̄ (s) = 1 − F (s)).
This dissertation does not concern itself with parameter estimation. For this reason,
the paramaters used in the programs are the ones commonly used in the literature.
The process parameters are p = 6, λ = 2 and µ = 0.5. The FORTRAN program
lberg.for in Appendix A has been used to obtain the results in Table 5.1. From Table
5.1, it can be seen that the results are excellent by any standards. Furthermore, the
results satisfy Theorem 4.1.1.
57
It can be seen by comparing Tables 5.1 and 5.2 that the smaller the step-size h,
the more accurate the results. The results in Table 5.2 (obtained with a step size 10
times that used for results in Table 5.1) are less accurate as is clear from comparing
the absolute percentage relative errors in the two tables. This is in line with our
expectation that the smaller the value of h, the better the results.
But, interestingly, when h decreases, say to 0.001, the results are less accurate,
contrary to what is expected. Table 5.3 shows the results after reducing the value of
h. The errors here are much higher than those in Table 5.1.
Then, by Kasozi and Paulsen (2005b), the survival probability φ(y) satisfies the
59
VIDE
Z y
0
(ry + p)φ (y) + λ (φ(y − s) − φ(y)) dF (s) = 0 (5.13)
0
Obviously, φ(y) = 0 for y < 0 (by (4.5)) and it is well known (by Theorem 4.1.1)
that lim φ(y) = 1. Integrating by parts on [0, y] brings (5.13) into a linear VIE of
y→∞
that is,
Z y
φ(y) = g(y) + K(y, s)φ(s)ds
0
p r+λF̄ (y−s)
where g(y) = ry+p
φ(0) and K(y, s) = ry+p
(with F̄ (s) = 1 − F (s)).
higher the value of the investment rate r, the lower the value of the ruin probability.
Thus, using r = 0.2 results in lower ruin probabilities than are obtained when the
surplus is invested at 10% (that is, when r = 0.1). Similarly, investing the surplus at
30% further reduces the ultimate ruin probabilities compared to the ones obtained
if the investment rate is 20% (as is clear from Table 5.4).
Next, assume that σP = σR = r = 0 and 0 < b < 1 (that is, the Cramér-Lundberg
model compounded by proportional reinsurance). Then (3.9) becomes
NP (t)
X
b
Y (t) = y + bpt − bSP,i
i=1
60
Table 5.4: Ruin probabilities for constant force of interest for p = 6, λ = 2 and
µ = 0.5
This VIDE reduces to the VIE of the second kind (4.14) with K(y, s) and α(y) given
by (4.17) when r = 0. For a given y, let ψb (y) be the calculated ruin probability when
a retention percentage b ∈ (0, 1] is used. Then Table 5.5 summarizes the results for
probabilities for the C-L model which are approximated by those in the third column
obtained without reinsurance (i.e., for b = 1). As can be seen by comparing with
Table 5.1 (for y = 0, y = 5 and y = 10) these are the same probabilities obtained
61
earlier. The ruin probabilities, obtained for small values of y, reduce very fast as
the retention percentage reduces (or as the insurance company insures more of its
portfolio of risks). This is to be expected, because the smaller the value of b, the
higher the reinsurance actually bought by the insurer and therefore the lower the
ruin probability.
It is clear from Table 5.5 that the optimal reinsurance percentage lies somewhere
between 20% and 30% (i.e., 0.2 ≤ b∗ ≤ 0.3), meaning that in order to minimize
its probability of ultimate ruin the insurance company should reinsure between 70%
and 80% of its portfolio of risks. According to Schmidli (2008), for exponential
claims the optimal choice of b that maximizes the adjustment coefficient R(b) is
R η
1
given by b = b ∧ 1, where b = 1 − θ 1 + 1+θ (θ and η are, respectively, the
√
safety loadings of the reinsurer and insurer). Because maximising the adjustment
coefficient yields the asymptotically best strategy, it is expected that the optimal b
will tend to bR . We assume in this study that the expected value premium principle
is used and that p = 6, θ = 6.5 and η = 5, thus fulfilling the net profit condition.
We also assume that λ = 2 and µ = 0.5 (so that λµ = 1). With these assumptions,
we have the asymptotically optimal bR = 0.315034. It is likely that b∗ is closer to
Table 5.5: Ruin probabilities for C-L model with proportional reinsurance for p = 6,
λ = 2, µ = 0.5, 0.2 ≤ b ≤ 1.0 and h = 0.01
Table 5.6: Ruin probabilities with exponentially distributed jumps for σP = 0.2
The exponential distribution, on which the above discussion has been based, is light-
tailed. This means that the probability of an extraordinary event is very small. For
this reason, it is safe to say that the insurer can withstand the risk from claims. To
investigate the effect of claim-size distributions more comprehensively, heavy-tailed
distributions also need to be studied.
can be studied is the Pareto distribution, which is often more appealing because it
allows the occurrence of extreme events, an important feature of claims in the real
world.
κ α
The distribution function for Pareto(α, κ), F (s) = 1 − κ+s
, will now be used to
show that the block-by-block method still works. The non-diffusion and diffusion
cases will both be investigated for the Cramér-Lundberg model with and without
64
ing the claim sizes are Pareto-distributed with α = 3 and κ = 2,the ruin probabilities
are computed for the Cramér-Lundberg model and for a constant force of interest.
For a given y, let the calculated ruin probability be given by ψh,r (y), where r ∈ [0, 1]
is the rate of return on investments. Then the results are summarized in Table 5.7
below. The second column shows results for the Cramér-Lundberg model while the
remaining three columns contain results when the Cramér-Lundberg model is com-
ble 5.4), the higher the investment rate, the lower the probability of ultimate ruin of
the insurance company.
ruin probability be given by ψh,r (y), where r ∈ [0, 1] is the rate of return on invest-
ments. Then the results are summarized in Table 5.8.
(in Appendix D) the asymptotic ruin probabilities are computed for the large claim
case with p = 6, θ = 6.5 and η = 5. The ruin probabilities are for the Cramér-
and retention level b ∈ (0, 1], let the calculated ruin probability be given by ψb (y).
The results for different values of b are summarized in Table 5.9 from which the
Table 5.9: Asymptotic ruin probabilities for large claims with proportional reinsur-
ance (for p = 6, θ = 6.5 and η = 5)
From Table 5.9 it can be concluded that the optimal retention percentage for Pareto-
distributed claims, assuming that p = 6, θ = 6.5 and η = 5, lies between 0.4 and 0.5
(i.e., 0.4 ≤ b ≤ 0.5). The actual optimal value b∗ can be obtained from
1 b 1
ψ b (y) ≈ y = 1.5
y
6.5b − 1.5 1 + b 6.5 − b
1+ b
3y
The right hand side is minimized for b = 6.5y−1.5
. Letting y → ∞ yields the optimal
3
bR = 6.5
= 0.461538. This is the asymptotically optimal proportional reinsurance
level b∗ . It is also evident from Table 5.9 that when y = 0, the ruin probabilities
increase as b reduces. This is because purchasing reinsurance when the insurance
company has no initial surplus only increases its indebtedness (e.g., to financial
institutions) and therefore seriously jeopardizes its chances of survival.
67
and σP = 0.2
Here S has the distribution (5.11) with α = 2, σP = 0.2 and σR = 0.2. The results,
obtained using the FORTRAN program sigma.for with appropriate adjustments,
are summarized in Table 5.10 and clearly show that, as for the exponential dis-
tribution, the ruin probabilities decreased exponentially fast with an increase in r
(compare with Table 5.6).
5.4 Discussion
happens, as can be seen from Table 5.4 in the previous section. For p = 6, λ = 0.2
and µ = 0.5, as the rate of return on investments increases, the ruin probability
reduces. This is true for both the small and large claim cases, as well as for the
diffusion and non-diffusion cases. This is because the return on investments is higher
if the interest rate is higher, thus the ruin probability is lower. It can therefore
be concluded that investments play a major role in helping insurance companies to
minimize their probabilities of ultimate ruin. The more of their surplus insurance
companies invest in risky and risk-free assets, the lower their ruin probabilities.
Insurers can never reach a point (in terms of investment rate) beyond which it is not
profitable to invest.
However, as stock prices become more volatile (that is, as σR increases), the ruin
probability also increases, and vice versa, as confirmed by Figure 5.1 below. Volatil-
ity is actually a measure of the riskiness of a stock. If the volatility of the stock price
increases but the expected rate of return of the stock stays the same, then the insurer
will find the reward for accepting the risk unattractive and would rather invest less
in stocks and invest more in bonds. Conversely, if the volatility of the stock price
decreases, then the insurer will receive the same return but with a lower risk and
will find that it makes economic sense to invest in the stock. The foregoing applies
to the exponential distribution but is also valid for other distributions (for example,
Pareto).
69
1
R=0.2
0.9
R=0.3
Ruin probability ( ) 0.8 R=0.4
0.7 R=0.5
0.6
0.5
0.4
0.3
0.2
0.1
0
0 5 10 15 20 25 30 35 40 45 50
Initial surplus (y)
Figure 5.1: Effect on ruin probability of changes in volatility of the stock price (σR )
From the findings on the impact of reinsurance on the ruin probability, it is evident
that for a given b ∈ (0, 1] the ruin probabilities keep reducing up to a given retention
level, after which they begin to increase. This is true for both the light-tailed (ex-
ponential) and heavy-tailed (Pareto) cases. This means that the optimal retention
level for proportional reinsurance lies somewhere around the point at which the ruin
probabilities begin to rise again after consistently falling with a reduction in b. This
is in line with our expectation that the ruin probabilities should keep reducing as the
retention percentage reduces and then start rising again after a certain b, giving an
indication of where the optimal retention percentage lies. The results from the pre-
vious section indicate that proportional reinsurance does have a positive impact on
the survival of insurance companies as it minimizes their ultimate ruin probabilities.
70
The findings show that insurance companies can reinsure optimally when b∗ =
0.315034 for small claims (for p = 6, λ = 0.2 and µ = 0.5) and when b∗ = 0.461538
for large claims (for p = 6, θ = 6.5, η = 5, where θ and η are, respectively, the safety
loadings of the reinsurer and insurer). This means that an insurance company should
reinsure 70% of its portfolio in the small claim case and about 55% of its risks in the
large claim case. The reason for this difference is that because large claims are also
extremal and therefore rare the company can afford to retain more of its large-scale
risks. Also, since small claims are more frequent the company should retain less of
them and cede a bigger chunk to a reinsurer.
5.5 Conclusion
In this chapter, a numerical method for solving the VIE (4.14) has been presented as
well as some numerical results based on this method both for small and large claims.
In this regard, results have been presented based on the exponential and Pareto
distributions for small and large claims, respectively. The chapter has also analyzed,
discussed and interpreted the findings based on each of the research objectives. The
next chapter draws conclusions and makes recommendations based on the research
findings and the analyses made.
CHAPTER SIX
CONCLUSION AND RECOMMENDATIONS
6.1 Introduction
This chapter draws conclusions based on the findings from the study. It also makes
recommendations which should be of benefit to insurance companies in terms of
measures they could take to minimize their ultimate ruin probabilities or to ensure
their survival. The chapter closes with an indication of open problems for future
research.
6.2 Conclusion
Findings from the study indicate that investment of the surplus plays an important
role in the survival of insurance companies as it significantly drives down the ultimate
6.3 Recommendations
Based on the findings from this study, it is recommended that insurance companies
should invest more of their surplus in risky and risk-free assets in order to enhance
their chances of survival. Furthermore, insurers should reinsure part of their portfo-
lios in order to minimize their probabilities of ultimate ruin. But, given the assump-
71
72
tions made in this study concerning the flow of premiums, insurers can only reinsure
optimally when b∗ = 0.315034 (small claims) or b∗ = 0.461538 (large claims).
to the model considered in this study. Another problem would involve extending
the model considered in this study by adding a compound Poisson process to the
investment generating process (as in Gjessing and Paulsen (1997) and Paulsen and
Gjessing (1997)) and then applying proportional reinsurance (or some other type of
reinsurance). Thus the investment generating process would be as in (3.6), that is,
NR (t)
X
R(t) = rt + σR WR (t) + SR,i , t≥0
i=1
numerical method used here would be appropriate for the VIE that would arise. The
task would involve identifying the appropriate forcing function and kernel to use in
the FORTRAN program.
73
REFERENCES
2, 295-320
Black, F. and Scholes M. (1973), The Pricing of Options and Corporate Lia-
Cai, J. and Xu, C. (2006), On the Decomposition of the Ruin Probability for
a Jump-Diffusion Surplus Process Compounded by a Geometric Brownian
Motion, North American Actuarial Journal, 2, 120-132
Castillo and Parrocha (2003), Stochastic Control Theory for Optimal Invest-
ment, Working Paper, Department of Actuarial Studies, University of New South
Wales
Volume, Stockholm
Dutang, C., Goulet, V. and Pigeon, M. (2009), Risk Theory Features of actuar
Furrer, H.J. and Schmidli, H. (1994), Exponential Inequalities for Ruin Prob-
abilities of Risk Processes Perturbed by Diffusion, Insurance: Mathematics
of Jyväskylä
76
Hipp, C. and Plum, M. (2003), Optimal Investment for Investors with State-
dependent Income, and for Insurers, Finance Stoch., 3, 299-321
sity of Iowa
Ma, J., Bai, L. and Liu, J. (2008), Minimizing the Probability of Ruin under
Interest Force, Applied Mathematical Sciences, 17, 843-851
Paulsen, J. (1998), Sharp Conditions for Certain Ruin in a Risk Process with
5, 416-434
Paulsen, J., Kasozi, J. and Steigen, A. (2005), A Numerical Method to Find the
Probability of Ultimate Ruin in the Classical Risk Model with Stochastic
Pitacco, E. (2004), ‘de Finetti, Bruno (1906-1985)’, in: Teugels, J. L. and Sundt,
B. (2004) (eds.), Encyclopedia of Actuarial Science, Vol. I, 421-423, Chichester, John
Rolski, T., Schmidli, H., Schmidt, V. and Teugels, J. L. (1999), Stochastic Pro-
Sundt, B. and Teugels, J. (1995), Ruin Estimates Under Interest Force, Insur-
sical Insurance Solvency Theory, ed. Cummins, J. D. and Derrig, R. A., Chapter 2,
49-151, Boston, MA, Kluwer Academic Publishers
the numerical solution of integral equations, Proc. Roy. Soc. London Ser. A,
224, 561-573
APPENDICES